Investigating the Performance of Collective Communications on SMP Clusters: A Case for MPI_Allgather
Feng Long Gu, Hyacinthe Nzigou M., Guilherme de Melo Baptista Domingues, Takeshi Nanri, Kazuaki J. Murakami, Theodore E. Simos, George Maroulis · AIP conference proceedings · 2007
Message‐passing interface (MPI) is a specification targeting parallel processing on Multiple instruction stream, multiple data stream (MIMD) architectures. Collective operation in MPI specification needs intensive communications. This paper reports an investigation on the performance of collective operations on PC clusters and multi‐core Symmetric MultiProcessor (SMP) computers. MPI_Allgather is one of the most relevant MPI collective communications. From the comparison between the different MPI Allgather algorithms (Ring and Pair‐wise) on different platforms, we describe how multi‐core systems can take the best advantage by exploiting difference between inter‐node and intra‐node communications.